Continuation-based learning algorithm for discrete-time cellular neural networks

The SGN-type nonlinearity of a standard discrete-time cellular neural network (DTCNN) is replaced by a smooth, sigmoidal nonlinearity with variable gain. Therefore, the resulting dynamical system is fully differentiable. Bounds on gain of the sigmoidal function are given, so that the new smooth system approximates the standard DTCNN within certain limits. A learning algorithm is proposed, which finds the template parameters for the standard DTCNN by gradually increasing the gain of the sigmoidal function.<
>